Text-to-Motion · SMPL Skeleton

SMPL Skeleton setting of the unified Text-to-Motion Leaderboard, evaluated on the HumanML3D official test split.

The public page contains the benchmark protocol, generated-method comparison charts, metric tables, and an all-case Three.js comparison covering all 4,042 selected-caption test cases and every released result row with retained output. The comparison decodes each method's motion_135 channels and animates a neutral SMPL body mesh with linear blend skinning; it is not a skeleton-only preview. Motion assets are fetched lazily from the public ZeyuLing/Motius-Leaderboard-Cases dataset so the Space itself remains lightweight. GT, paper-only, and explicitly marked calibration rows are excluded from all generated-method rankings. Internal result paths and debugging notes stay in the repository documentation and experiment logs.